RSI Arena: 10 AI agents self-train models; human feedback lifts 8 of 10
my_cat_can_code · x · 2026-11-06
Bake AI, with Hugging Face, NVIDIA, Scale AI, Stanford, MIT and others, ran RSI Arena: 10 frontier-model agents each got the same NVIDIA Nemotron 3.5 Lightning 30B-A3B base and a shared 64-GPU cluster to pick data, write training code, and train their own assistants — 13,000 GPU-hours total, 20 benchmarks, 3,129 blind human votes across two rounds.
- The strongest agent (GPT-6 Astra) closed a median 76% of the base-to-Instruct gap on 12 capability benchmarks; its human-preference rating near-matched NVIDIA Instruct after length/format adjustment.
- In stage 2, agents got their round-1 feedback and 3 more days; 8 of 10 improved style-adjusted ratings — but benchmark gains didn't track preference gains.
- Takeaway: AI can do more of its own development, but human feedback still defines what "improvement" means.
Related event: RSI Arena: AI Agents Autonomously Fine-Tune Models with Measurable Gains(2 posts)→
More from Research
- Shadaj Lohia presents Hydro: chasing zero-cost correctness beyond C++, Zig and Go — ShadajL · 2026-11-06
- Lighting a Christmas tree with Stein Variational Gradient Descent — PTenigma · 2026-11-06
- Alex Wang's Perplexity reading group slides reframe Transformer design through associative memory — andrewgwils · 2026-11-06
- Peking University-led team open-sources 50K-hour recipe for embodied foundation models — DAGroup-PKU · 2026-11-06
- NPCBench: a clinical apprenticeship benchmark for longitudinal multimodal cancer care reasoning — InfiX-ai · 2026-11-06
- 10 open-source GitHub repos that research topics and write cited reports — victor_explore · 2026-11-06